Collision Avoidance Navigation with Radar and Spiking Reinforcement Learning

Collision avoidance is a safety-critical function in autonomous navigation. Radar is seen as a good sensor for this because of its robustness in severe weather and poor lighting conditions. Several deep reinforced learned approaches exist where an agent tries to reach a goal using collision avoidance as a constraint. However, these solutions make use of neural networks of substantial size and thus consume much of the available processing power. This conflicts with an important requirement within autonomous systems - energy efficiency. Spiking Neural Networks (SNN) could provide a solution to this problem. They use event-based communication, which is fast, efficient, and sparse by design.This paper proposes a novel obstacle avoidance technique using a low compute method to encode radar range azimuth spectra into spike trains and a very small SNN controller. The SNN is trained in simulation with a biologically inspired reinforcement learning technique Reward Modulated Spike Timing Dependent Plasticity. Experiments on both simulated and real-life environments show that, when using a single FMCW radar and a very small SNN of only 414 neurons, our approach is capable of performing collision avoidance while trying to reach a point goal with an average collision rate of 5% over a variety of scenarios.